Abstract: Strategies from convex optimization are widely utilized as setting up blocks for deep finding out algorithms. Nevertheless, the reasons for their empirical good results are unclear, considering the fact that fashionable convolutional networks (convnets), incorporating rectifier models and max-pooling, are neither smooth nor convex. Conventional ensures for that reason do not use. This paper presents the 1st convergence rates for gradient descent on rectifier convnets. The evidence makes use of the particular structure of rectifier networks which is made up in binary energetic/inactive gates applied on top of an underlying linear network.